---

   

Sihui Ji¹ Xi Chen¹ Xin Tao² Pengfei Wan² Hengshuang Zhao^1✉

¹The University of Hong Kong    ²Kling Team, Kuaishou Technology
^✉Corresponding author

📌 Updates

• [2025.10.16]: Release the Project Page and the Arxiv version.

📷 Introduction

TL;DR: We propose PhysMaster, which captures physical knowledge as a representation for guiding video generation models to enhance their physics-awareness. Specifically, PhysMaster is based on the image-to-video task where the model is expected to predict physically plausible dynamics from the input image. We devise PhysEncoder to encode such physical representation as an extra condition, and adopt a top-down optimization strategy for finetuning PhysEncoder based on the physical plausibility of the final generated videos using reinforcement learning (RL). Experiment results demonstrate strong performance of our model on both specialized proxy tasks and general open-world scenarios.

🌄 Gallery

Comparisons with Previous Works

compare.mp4

Ablation Analysis

ablation.mp4

⚖ License

This repository is released under the Apache-2.0 license as found in the LICENSE file.

📚 Citation

If you find this codebase useful for your research, please use the following entry.

@article{Ji2025physmaster,
  title={{PhysMaster: Mastering Physical Representation for Video Generation via Reinforcement Learning}},
  author={Ji, Sihui and Chen, Xi and Tao, Xin and Wan, Pengfei and Zhao, Hengshuang},
  journal={arXiv preprint arXiv:2510.13809},
  year={2025}
}